-
Notifications
You must be signed in to change notification settings - Fork 1
/
model_SENet.py
283 lines (247 loc) · 13.7 KB
/
model_SENet.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
import tensorflow as tf
import numpy as np
from Basic_function import conv_layer,Max_pooling,Fully_connected,Relu,SqueezeExcitation,Avg_pooling,Dropout,Global_Average_Pooling
class SENet:
def __init__(self,input_shape,dataset,batch_size,epoch,learning_rate):
self.input_shape=input_shape
self.x_input = input_shape[0]
self.y_input = input_shape[1]
self.channel = input_shape[2]
self.dataset=dataset
self.batch_size=batch_size
self.epoch=epoch
self.learning_rate=learning_rate
self.checkpoint_path='./save/SENet'+dataset+'Epoch'+str(self.epoch)+'.ckpt'
if dataset=="Cifar":
self.output_num=10
self.Train_data="./Data_to_record/CifarRecord/train.tfrecords"
self.Test_data = "./Data_to_record/CifarRecord/test.tfrecords"
self.Validation_data = "./Data_to_record/CifarRecord/validation.tfrecords"
elif dataset=="BlindSpot":
self.output_num=1
self.datasource = "./"
else:
print ("There is no such dataset. It should be 'Cifar' or 'BlindSpot'.")
tf.set_random_seed(1337)
np.random.seed(1337)
def Data_input(self,data_path,Epoch=1):
feature = {'example': tf.FixedLenFeature([], tf.string),
'label': tf.FixedLenFeature([], tf.int64)}
# Create a list of filenames and pass it to a queue
filename_queue = tf.train.string_input_producer([data_path], num_epochs=Epoch)
# Define a reader and read the next record
reader = tf.TFRecordReader()
_, serialized_example = reader.read(filename_queue)
# Decode the record read by the reader
features = tf.parse_single_example(serialized_example, features=feature)
# Convert the image data from string back to the numbers
image = tf.decode_raw(features['example'], tf.float32)
# Cast label data into int32
label = tf.cast(features['label'], tf.int32)
# Reshape image data into the original shape
image = tf.reshape(image, self.input_shape)
# Any preprocessing here ...
# Creates batches by randomly shuffling tensors
images, labels = tf.train.shuffle_batch([image, label], batch_size=self.batch_size, capacity=100, num_threads=1,
min_after_dequeue=10)
return (images,labels)
def model_preprocess(self):
Mean_RGB = [0, 0, 0]
with tf.Session() as sess:
images, labels = self.Data_input(data_path=self.Train_data, Epoch=1)
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess.run(init_op)
# Create a coordinator and run all QueueRunner objects
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
count=0.0
for j in range(100000):
try:
img, lbl = sess.run((images, labels))
lbl = np.reshape(lbl, (self.batch_size, 1))
except:
break
for m in range(self.batch_size):
assert np.shape(img[m])==(32,32,3)
count=count+1
Mean_RGB[0] = Mean_RGB[0] + np.mean(img[m][:, :, 0])
Mean_RGB[1] = Mean_RGB[1] + np.mean(img[m][:, :, 1])
Mean_RGB[2] = Mean_RGB[2] + np.mean(img[m][:, :, 2])
Mean_RGB[0]=Mean_RGB[0]/count
Mean_RGB[1] = Mean_RGB[1] / count
Mean_RGB[2] = Mean_RGB[2] / count
coord.request_stop()
# Wait for threads to stop
coord.join(threads)
return Mean_RGB
def model_structure(self):
self.img = tf.placeholder(tf.float32, shape=(None, self.x_input, self.y_input, self.channel))
self.label = tf.placeholder(tf.int32, shape=(None, 1))
self.keep_prob = tf.placeholder(tf.float32)
if self.dataset=='Cifar':
self.conv_1 = conv_layer(self.img, filter=64, kernel=[3, 3], stride=1, padding='SAME', layer_name="conv_1")
self.Relu_1=Relu(self.conv_1)
self.Pool_1 = Max_pooling(self.Relu_1, pool_size=[2, 2], stride=2, padding='VALID')
if self.dataset=='BlindSpot':
self.conv_1 = conv_layer(self.img, filter=64, kernel=[7, 7], stride=2, padding='SAME', layer_name="conv_1")
self.Relu_1 = Relu(self.conv_1)
self.Pool_1 = Max_pooling(self.Relu_1, pool_size=[2, 2], stride=2, padding='VALID')
self.conv_2 = conv_layer(self.Pool_1, filter=128, kernel=[3, 3], stride=1, padding='SAME', layer_name="conv_2")
self.SElayer_1 = SqueezeExcitation(self.conv_2, input_channel=128, R=8, batch_size=self.batch_size,se_name="SE_1")
self.Relu_2 = Relu(self.SElayer_1)
self.conv_3 = conv_layer(self.Relu_2, filter=128, kernel=[3, 3], stride=1, padding='SAME', layer_name="conv_3")
self.SElayer_2=SqueezeExcitation(self.conv_3, input_channel=128, R=8, batch_size=self.batch_size,se_name="SE_2")
self.Relu_3 = Relu(self.SElayer_2)
self.conv_4 = conv_layer(self.Relu_3, filter=256, kernel=[3, 3], stride=1, padding='SAME', layer_name="conv_4")
self.SElayer_3 = SqueezeExcitation(self.conv_4, input_channel=256, R=8, batch_size=self.batch_size,se_name="SE_3")
self.Relu_4 = Relu(self.SElayer_3)
self.conv_5 = conv_layer(self.Relu_4, filter=256, kernel=[3, 3], stride=1, padding='SAME', layer_name="conv_5")
self.SElayer_4 = SqueezeExcitation(self.conv_5, input_channel=256, R=8, batch_size=self.batch_size,se_name="SE_4")
self.Relu_5 = Relu(self.SElayer_4)
self.Pool_2 = Avg_pooling(self.Relu_5, pool_size=[2, 2], stride=2, padding='VALID')
self.flatten = tf.layers.flatten(self.Pool_2)
self.FC_0 = Fully_connected(self.flatten, out_num=1024, layer_name='fc_0')
self.Relu_0 = Relu(self.FC_0)
self.FC_1 = Fully_connected(self.Relu_0, out_num=128, layer_name='fc_1')
self.Relu_6 = Relu(self.FC_1)
self.Dropout_1 = Dropout(self.Relu_6, self.keep_prob)
self.FC_2 = Fully_connected(self.Relu_6, out_num=32, layer_name='fc_2')
self.Relu_7 = Relu(self.FC_2)
#self.Dropout_2 = Dropout(self.Relu_7, self.keep_prob)
self.FC_3 = Fully_connected(self.Relu_7, out_num=self.output_num, layer_name='fc_3')
if self.output_num == 1:
self.prob = tf.nn.sigmoid(self.FC_3, name="prob")
else:
self.prob = tf.nn.softmax(self.FC_3, name="prob")
return self.prob
def fit(self):
tf.reset_default_graph()
self.RGB_mean_value =self.model_preprocess()
Output_Probability= self.model_structure()
if self.output_num==1:
cross_entropy=-tf.reduce_mean(tf.multiply(self.label,tf.log(Output_Probability+0.0001))+tf.multiply((1.0-Output_Probability),tf.log(1.0-self.prob+0.0001)))
else:
self.one_hot=tf.one_hot(indices=self.label,depth=self.output_num)
self.one_hot=tf.squeeze(self.one_hot,[1])
cross_entropy=-tf.reduce_mean(tf.reduce_mean(tf.multiply(tf.log(self.prob+0.0001),self.one_hot),axis=1),axis=0)
Learning_rate=tf.placeholder(tf.float32,None)
trainer=tf.train.AdamOptimizer(Learning_rate)
gvs=trainer.compute_gradients(cross_entropy)
train_step = trainer.apply_gradients(gvs)
'''
def ClipGradient(grad):
if grad is None:
return grad
return tf.clip_by_value(grad,-10,10)
clip_gradient=[]
for grad,var in gvs:
clip_gradient.append((ClipGradient(grad),var))
train_step=trainer.apply_gradients(clip_gradient)
'''
with tf.Session() as sess:
train_accuracy = 0
Train_Step = 0
images, labels = self.Data_input(data_path=self.Train_data, Epoch=self.epoch)
vali_images, vali_labels = self.Data_input(data_path=self.Validation_data, Epoch=200)
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess.run(init_op)
# Create a coordinator and run all QueueRunner objects
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
def Validation(step):
validation_accuracy = 0
validation_Step = 0
for j in range(100):
try:
img, lbl = sess.run([vali_images, vali_labels])
img[:, :, :, 0] = img[:, :, :, 0] - self.RGB_mean_value[0]
img[:, :, :, 1] = img[:, :, :, 1] - self.RGB_mean_value[1]
img[:, :, :, 2] = img[:, :, :, 2] - self.RGB_mean_value[2]
lbl = np.reshape(lbl, (self.batch_size, 1))
except:
break
validation_Step = validation_Step + 1
Probability = Output_Probability.eval(
feed_dict={self.img: img, self.label: lbl, self.keep_prob: 1.0})
Probability = np.asarray(Probability)
if self.output_num == 1:
assert np.shape(Probability) == (self.batch_size, 1)
validation_accuracy = validation_accuracy + np.mean(
(Probability > 0.5) * lbl + (Probability <= 0.5) * (1 - lbl))
else:
assert np.shape(Probability) == (self.batch_size, self.output_num)
validation_accuracy = validation_accuracy + +np.mean(
np.reshape(np.argmax(Probability, axis=1), (self.batch_size, 1)) == lbl)
print("Step %d validation accuracy is: %f" % (step, validation_accuracy / validation_Step))
for j in range(10000000):
try:
img, lbl = sess.run((images, labels))
img[:, :, :, 0] = img[:, :, :, 0] - self.RGB_mean_value[0]
img[:, :, :, 1] = img[:, :, :, 1] - self.RGB_mean_value[1]
img[:, :, :, 2] = img[:, :, :, 2] - self.RGB_mean_value[2]
lbl=np.reshape(lbl,(self.batch_size,1))
except:
break
Train_Step=Train_Step+1
if j%100==0 and j>5000:
self.learning_rate=self.learning_rate*0.99
train_step.run(feed_dict={self.img:img,self.label:lbl,Learning_rate:self.learning_rate,self.keep_prob:0.6})
Probability=Output_Probability.eval(feed_dict={self.img:img,self.label:lbl,self.keep_prob:1.0})
Probability=np.asarray(Probability)
if self.output_num==1:
assert np.shape(Probability)==(self.batch_size,1)
train_accuracy=train_accuracy+np.mean((Probability>0.5)*lbl+(Probability<=0.5)*(1-lbl))
else:
assert np.shape(Probability)==(self.batch_size,self.output_num)
train_accuracy=train_accuracy+np.mean(np.reshape(np.argmax(Probability,axis=1),(self.batch_size,1))==lbl)
if j%100==0 and j>0:
print ("Step %d accuracy is : %f"%(int(j),train_accuracy/j))
if j%5000==0 and j>1:
Validation(j)
print ("Epoch %d training accuracy is: %f"%(self.epoch,train_accuracy/Train_Step))
coord.request_stop()
# Wait for threads to stop
coord.join(threads)
saver=tf.train.Saver()
saver.save(sess,self.checkpoint_path)
def test_check(self):
tf.reset_default_graph()
self.RGB_mean_value = self.model_preprocess()
Output_Probability = self.model_structure()
with tf.Session() as sess:
test_accuracy = 0
Test_Step = 0
test_images, test_labels = self.Data_input(data_path=self.Test_data, Epoch=1)
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess.run(init_op)
saver=tf.train.Saver()
saver.restore(sess,self.checkpoint_path)
# Create a coordinator and run all QueueRunner objects
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
for j in range(1000000):
try:
img, lbl = sess.run([test_images, test_labels])
img[:, :, :, 0] = img[:, :, :, 0] - self.RGB_mean_value[0]
img[:, :, :, 1] = img[:, :, :, 1] - self.RGB_mean_value[1]
img[:, :, :, 2] = img[:, :, :, 2] - self.RGB_mean_value[2]
lbl = np.reshape(lbl, (self.batch_size, 1))
except:
break
Test_Step = Test_Step + 1
Probability = Output_Probability.eval(feed_dict={self.img: img, self.label: lbl,self.keep_prob:1.0})
Probability = np.asarray(Probability)
if self.output_num == 1:
assert np.shape(Probability) == (self.batch_size, 1)
test_accuracy = test_accuracy + np.mean(
(Probability > 0.5) * lbl + (Probability <= 0.5) * (1 - lbl))
else:
assert np.shape(Probability) == (self.batch_size, self.output_num)
test_accuracy = test_accuracy + +np.mean(
np.reshape(np.argmax(Probability, axis=1), (self.batch_size, 1)) == lbl)
print("Epoch %d test accuracy is: %f" % (self.epoch, test_accuracy / Test_Step))
coord.request_stop()
# Wait for threads to stop
coord.join(threads)
def predict(self):
pass